# Ask Paper > Answer questions about a paper in the vault, using note first and arXiv HTML as fallback - Skill: `majiayu000/ask-paper` (Agent Skill, multi-file: 2 files) - Install (CLI): `npx skillmds add majiayu000/ask-paper` - Raw SKILL.md: https://api.skillmd.com/api/skills/majiayu000/ask-paper/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Web & Frontend - Author: majiayu000 (https://skillmd.com/u/majiayu000) - Updated: 2026-09-09 - Page: https://skillmd.com/skills/majiayu000/ask-paper --- ## ❓ Ask Paper Skill Answer a question about a specific paper in the vault. **Input format:** `/ask_paper | ` If no `|` separator is present, treat the entire argument as combined input and infer the paper title and question from context. --- ## CRITICAL RULES - **Read the paper note first** — always attempt to answer from the note - **Only fetch arXiv HTML if the note is insufficient** — avoid unnecessary web requests - **Never hallucinate** — if the answer isn't in the note or the paper, say so - **Single paper scope** — answer questions about one paper at a time - **Project context allowed** — may reference linked projects from the note to give research-relevant answers - **Update the note** when arXiv HTML provides significant new information relevant to the question --- ## WORKFLOW ### 1. Resolve paper identity **Fuzzy matching:** - Search `Reading/` directory for files matching the provided title/keyword - Use glob pattern: `Reading/**.md` - If multiple matches, pick the closest match or ask user to disambiguate - If no match, inform user the paper is not in the vault **Validation:** - Confirm the matched file exists in `Reading/` - Extract paper title from file name ### 2. Read the paper note - Read the full paper note from `Reading/.md` - Extract all sections and content - Extract the `url` field from YAML frontmatter (needed for fallback) - Extract linked projects from "Linked Research Projects" section ### 3. Attempt to answer from the note **Evaluate confidence:** - Can the question be fully answered using information in the note? - Is the answer specific enough, or would it benefit from more detail? **Confidence levels:** - **HIGH**: Answer is clearly and specifically covered in the note → answer directly, skip arXiv fetch - **MEDIUM**: Answer is partially covered but lacks detail → answer from note, then fetch arXiv for enrichment - **LOW**: Answer is not covered in the note at all → must fetch arXiv HTML **If HIGH confidence:** Answer the question using note content. Done. **If MEDIUM or LOW:** Proceed to step 4. ### 4. Fetch arXiv HTML (fallback) **Get the URL:** - Use the `url` field from the paper note's YAML frontmatter - The URL should be an arXiv HTML link (e.g., `https://arxiv.org/html/[paper-id]`) - If URL is missing or `none`, inform user and attempt to find it via paper title search **Fetch with targeted prompt:** - Use WebFetch with a prompt specifically targeting the user's question - Do NOT extract the entire paper — focus the prompt on the relevant section - Example prompt: "Find information about [specific topic from question] in this paper. Look in methods, results, and discussion sections." **Token efficiency:** - Craft the WebFetch prompt to extract only the information needed to answer the question - Avoid generic "extract everything" prompts - If the question is about methods → target methods section - If the question is about results → target results/experiments section - If the question is about a specific concept → target the relevant section + definitions ### 5. Answer the question **Format:** - Bullet points preferred (consistent with vault style) - Cite specific sections/figures/tables when possible - If project context is relevant, mention which linked projects the answer connects to - Be concise — answer the question, don't summarize the whole paper **Source attribution:** - If answered from note only: no attribution needed - If answered from arXiv HTML: note that this information was fetched from the paper directly ### 6. Update the paper note (when arXiv HTML was fetched) **When to update:** - Only when arXiv HTML provided information NOT already in the note - Only when the new information is significant (not trivial details) - Only add information relevant to the question and the note's existing structure **How to update:** - Identify the most appropriate existing section in the note for the new information - Add bullet points under the relevant subsection - Do NOT restructure or rewrite existing content - Do NOT add new sections — use existing template sections only - Keep additions concise (1-3 bullets max) **What to add:** - Specific technical details that enrich an existing section - Key results or numbers not captured in the note - Method details that clarify the approach - Important caveats or nuances **What NOT to add:** - Information already in the note (even if phrased differently) - Tangential details not related to the question - Full section rewrites **After updating:** - Briefly inform the user what was added to the note and where --- ## HANDLING EDGE CASES ### Paper not in vault - Inform user: "Paper not found in Reading/. Would you like me to read it first using /read_paper?" ### No arXiv URL in frontmatter - Check if `url` field is `none` or missing - Inform user the URL is not available - Suggest adding the URL to the frontmatter ### Question requires cross-referencing - If the question implicitly compares to another paper, answer only for the target paper - Mention: "For comparison with [other paper], you could run /ask_paper on that paper as well." ### Question about linked projects - Read the linked project files if needed to provide project-relevant context - Only read projects listed in the "Linked Research Projects" section - Keep project context brief — focus on the paper question --- ## TOKEN EFFICIENCY **Optimizations:** - Always try note first (zero web cost) - Targeted WebFetch prompts (skip irrelevant paper sections) - Read only the specific project files if project context is needed - Single WebFetch call in most cases **Expected cost:** - Note-only answer: ~50 tokens (just file read) - ArXiv fallback: ~100-200 tokens (one targeted WebFetch) --- ## OUTPUT RULES - Bullet points preferred - Concise, direct answers - Cite sections/figures/tables when referencing arXiv content - Mention source (note vs arXiv) only when arXiv was fetched - If note was updated, report what was added and where - No essays — answer the question and stop --- ## SELF-CHECK ✅ Did I search Reading/ for the paper using fuzzy matching? ✅ Did I read the full paper note before attempting to answer? ✅ Did I evaluate confidence before fetching arXiv HTML? ✅ Did I use a targeted WebFetch prompt (not generic)? ✅ Did I answer concisely with bullet points? ✅ If I fetched arXiv: did I update the note with significant new information? ✅ If I updated the note: did I inform the user what was added? ✅ Did I reference linked projects when relevant to the question?